Circulating metabolomic signatures of ultra-processed and minimally processed food intake and risk of cardiovascular disease and type 2 diabetes: 20-year follow-up in the Whitehall II cohort Authors: Mengmei E. Wang, Annie Britton, Michail Katsoulis, Kiara Chang, Eszter P. Vamos, Clare H. Llewellyn Published: 2026 ABSTRACT BackgroundHigher intakes of ultra-processed food (UPF) and minimally processed food (MPF) are linked to increased and decreased cardiometabolic risks, respectively, though mechanisms underlying these associations remain less understood. We aimed to identify metabolomic signatures associated with UPF and MPF intakes, examine their prospective associations with cardiovascular disease (CVD) and type 2 diabetes (T2D), and quantify their mediation of processing-disease associations.MethodsIn the Whitehall II cohort, UPF and MPF intakes in 6010 participants were derived using the Nova classification from 127-item food frequency questionnaires administered at phases 3 and 5 (1991–1994, 1997–1999), averaged across phases. Fasting plasma samples were analysed for 233 metabolites using nuclear magnetic resonance (NMR) spectroscopy. Metabolomic signatures for UPF and MPF intake were derived using elastic net regression. Cox models examined associations between metabolomic signatures and incident CVD and T2D. Mediation analyses quantified the proportion explained by metabolomic signatures.ResultsWe identified metabolomic signatures for UPF (34 metabolites) and MPF (50 metabolites). The UPF signature correlated positively with pro-inflammatory and atherogenic metabolites and inversely with omega-3 fatty acids and high-density lipoproteins; the MPF signature showed the opposite pattern. 1094 CVD and 853 T2D events occurred over 20 and 18 years median follow-up, respectively. Comparing the highest to lowest tertile, higher UPF signature scores were associated with greater risk of (hazard ratio [HR] = 1.19; 95% confidence interval [CI] 1.02–1.39) and T2D (HR = 1.20; 95% CI 1.03–1.42), whereas higher MPF signature scores were inversely associated with these outcomes (CVD: HR = 0.81; 95% CI 0.69–0.94; T2D: HR = 0.60; 95% CI 0.48–0.73). Associations between the signatures and outcomes were robust to additional adjustment for UPF or MPF intake. The signatures mediated 42.8–51.1% of processing-disease associations (51.1% for UPF-CVD, 46.6% for MPF-CVD, 42.8% for UPF-T2D, and 47.8% for MPF-T2D; all p < 0.05).ConclusionsMetabolomic signatures of UPF and MPF intake showed opposing associations with CVD and T2D risk, partially mediating these processing–disease associations. These findings suggest biological pathways linking UPF intake to adverse cardiometabolic outcomes, though the unexplained proportion indicates additional mechanisms may contribute. Background Cardiovascular disease (CVD) and type 2 diabetes (T2D) account for over 20 million deaths annually and affect more than one billion people worldwide [1]. Each poses a major public health and economic challenge, contributing significantly to healthcare costs and loss of productivity [1, 2]. These conditions frequently co-occur and share underlying pathophysiological mechanisms, including chronic inflammation, insulin resistance, and dyslipidemia [3]. While dietary factors represent key modifiable determinants of cardiometabolic risk, the rapid global dietary transition toward industrial food processing has coincided with rising disease incidence [4]. In particular, ultra-processed foods (UPF) have been in the spotlight in recent years. UPF, defined by the Nova classification, are industrial formulations made mostly or entirely from substances extracted or derived from foods, together with cosmetic additives, and typically contain little or no intact whole food [5]. UPF now constitutes 50–70% of energy intake in the UK, 14–44% across the rest of Europe and are rapidly expanding in low- and middle-income countries [6, 7]. An umbrella review reported that higher UPF consumption was associated with increased risks of T2D and CVD [8]. UPF tend to be more energy dense and higher in added sugars, saturated fat, and sodium than non-UPF, which may contribute to adverse cardiometabolic risk through established pathways including obesity, insulin resistance, and hypertension [8–13]. However, the review reported that the associations remained robust in most studies after adjustment for dietary composition and other potential confounders [14–16]. Processing-specific factors may explain these independent effects: techniques such as high-temperature treatment, chemical modification, and the use of novel food additives can disrupt metabolic regulation through pathways not captured by nutrient composition alone [13, 17, 18]. Despite growing epidemiological evidence, the molecular mechanisms linking UPF intake to cardiometabolic disease remain poorly understood. Observational studies have shown that associations between UPF intake and adverse health outcomes persist after adjustment for overall dietary quality and macronutrient composition, suggesting that health effects may not be fully explained by nutritional composition alone [19]. This gap highlights the critical need for mechanistic insights into how food processing influences human metabolism beyond traditional nutritional parameters. Metabolomics profiling has emerged as a promising approach for investigating UPF-cardiometabolic health relationships, offering potential insights into the biological mechanisms linking high UPF intake with poorer health outcomes [20]. Recent controlled feeding trials have shown that diets composed of 80 versus 0% energy from UPF produced differences in plasma and urine metabolites [21], supporting the use of metabolomics to characterise UPF dietary patterns. Elucidating UPF-associated metabolomic signatures may help clarify the mechanisms through which food processing levels are associated with cardiometabolic health. Concurrently, growing interest in the health benefits of MPF warrants investigation of the potential favourable biological pathways, which remain insufficiently characterised. To date, few studies have directly compared the metabolomic signatures of UPF and MPF intake. To our knowledge, only one study has analysed both in parallel [22], and most investigations have been constrained by narrow metabolite coverage, cross-sectional designs, or single time-point dietary assessments [23, 24]. Critically, no prospective study has investigated how metabolomic signatures of UPF and MPF relate to long-term cardiometabolic disease risk [22, 25]. This limits our ability to determine whether observed health risks are attributable to adverse metabolite profiles associated with UPF intake or to the absence of favourable metabolite profiles associated with MPF intake, or both. We therefore leveraged data from a large, well-characterised population-based cohort with repeated dietary assessments and extended follow-up, to: (1) identify metabolomic signatures associated with UPF and MPF intake; (2) examine their prospective associations with incident CVD and T2D; and (3) quantify the proportion of processing-disease associations mediated by these metabolomic signatures. Methods Study population Data were from the Whitehall II cohort study, established in 1985 using 10,308 British civil servants (67% men) aged 35–55 years recruited from 20 London-based government departments [26]. Follow-up clinical examinations have occurred approximately every five years, with the latest examination in 2019–2022 (phase 13). Participants provided written informed consent, and ethical approval was obtained from the University College London Hospital Committee on the Ethics of Human Research (reference number 85/0938). In this study, we included participants who attended phase 5 clinical examinations (n = 8628) with dietary intake data from either phase 3 (1991–1994), phase 5 (1997–1999), or both. As metabolomics data were measured at phase 5 only, dietary intakes from phases 3 and 5 were averaged to capture cumulative dietary exposure and to establish temporal precedence of dietary exposure before metabolite measurement. The study timeline and data collection phases are illustrated in Additional file 1: Fig. S1. Exclusions included: missing dietary data at both phases (n = 1374), implausible energy intake (<500 or >3500 kcal for women; <800 or >4000 kcal for men; n = 470), death or loss to follow-up before phase 5 (n = 124), and non-participation in phase 5 examinations (n = 379), yielding 6281 participants with metabolite and dietary data. Participants with existing CVD (n = 170) or T2D (n = 101) at phase 5 (baseline) were excluded, yielding a final analytical sample of 6010 participants (Additional file 1: Fig. S2). We compared key baseline characteristics between participants included in the analytic sample and those excluded (Additional file 2: Table S1). Characteristics were broadly similar between groups. Measurement of metabolomics Whitehall II utilised the Nightingale Health NMR (nuclear magnetic resonance) metabolomics platform (Helsinki, Finland) for the analysis of metabolites [27]. The metabolite panel comprised 233 metabolites measured by NMR spectroscopy, including: lipoprotein subclasses (very-low-density lipoprotein (VLDL), intermediate-density lipoprotein (IDL), low-density lipoprotein (LDL), and high-density lipoprotein (HDL)); cholesterol and lipid measures (total and free cholesterol, triglycerides, phospholipids); fatty acids (including omega-3, omega-6, and saturation ratios); apolipoproteins (A1 and B); glycolysis-related metabolites (glucose, lactate, pyruvate, citrate); amino acids; ketone bodies; fluid balance markers (albumin, creatinine); and inflammation-related markers (e.g., glycoprotein acetyls). The full list of 233 metabolites is provided in Additional file 2: Table S2. All metabolite concentrations were log-transformed and standardised to reduce skewness. Dietary assessment Dietary intake was assessed at phase 3 and phase 5 using a validated 127-item semi-quantitative food frequency questionnaire (FFQ); detailed methodology has been reported elsewhere [28]. Two researchers independently classified each FFQ item according to Nova, with discrepancies resolved by consensus. Items potentially spanning multiple processing levels were assigned to the category corresponding to the most commonly consumed form in the UK [5]; where ambiguity remained, items were conservatively assigned to lower Nova categories, favouring attenuation of observed associations. The complete Nova classification of the Whitehall II FFQ has been published previously [12, 29]. Foods were also assigned to subcategories based on food type: UPF subcategories included savoury snacks, beverages, sauces, dairy products, UPF meats, sweets, cereals, breads and ready-to-eat meals; MPF subcategories comprised fresh fruits and vegetables, whole grains, legumes and beans, dairy products, meats and poultry, fish, and beverages (water, unsweetened tea and coffee), consistent with prior operationalisations in prospective cohort studies [30, 31]. UPF and MPF intakes were calculated as percentage of total daily food weight (grams) rather than energy, to capture non-nutritional processing factors, particularly in low- or zero-calorie ultra-processed products such as non-nutritively sweetened beverages [32]. Details of food item classification according to the Nova system, along with the categorisation of UPF and MPF subgroups, are provided in Additional file 2: Tables S3–S4. Covariates assessment In line with prior literature, we adjusted for covariates measured at phase 5 (1997–1999), including sociodemographic characteristics, health behaviours, dietary indicators, and clinical measures, obtained via self-reported questionnaires and standardised clinical examinations [26]. Socio-demographic factors included age, sex, marital status (married/cohabiting vs. other), ethnicity (white vs. non-white), educational attainment (under secondary, secondary, or university level), and socioeconomic status (SES) based on civil service employment grade (low [clerical/support], intermediate [professional/executive], or high [administrative]). Behavioural factors included smoking status (current, former, or never), physical activity levels (active [>2.5 h/week of moderate physical activity or >1 h/week of vigorous physical activity], inactive [<1 h/week of both moderate and vigorous physical activity], or moderately active [neither active nor inactive]), and alcohol intake (abstainers [none in the past week], moderate [1–14 units/week], or heavy [≥15 units/week], where 1 unit = 8 g alcohol) [33, 34]. Dietary factors comprised total energy intake (in kilocalories) as well as sodium (mg/day), total fat (g/day), sugar intake (g/day), fibre (g/day) and Alternative Healthy Eating Index-2010 (AHEI-2010). Clinical factors were considered as follows: family history of CVD and T2D, body mass index (BMI) was calculated as weight in kilograms divided by height in metres squared (kg/m²) and modelled as a continuous variable. Height and weight were obtained in clinic by trained staff; full measurement details are reported elsewhere [35]. Hypertension was defined as either systolic blood pressure ≥140 mmHg, diastolic blood pressure ≥90 mmHg, or the use of antihypertensive medication. Dyslipidaemia was defined as meeting any one of the following criteria: total cholesterol >6.0 mmol/L; triglycerides >1.7 mmol/L; HDL-cholesterol <1.0 mmol/L; LDL-cholesterol >4.0 mmol/L; or receiving lipid-lowering medication [36]. Outcome ascertainment The primary outcomes were incident CVD and T2D. CVD cases were identified through Hospital Episode Statistics (HES) and mortality registers using International Classification of Diseases (ICD) 9/10 codes, with follow-up to October 2019. T2D cases were ascertained from clinical assessments and HES linkage (to March 2017) based on: (1) fasting glucose ≥7.0 mmol/L; (2) glucose-lowering medication use; (3) oral glucose tolerance test (OGTT) showing fasting glucose ≥7.0 mmol/L or 2-h glucose ≥11.1 mmol/L; (4) HbA1c ≥6.5%; or (5) self-reported or HES-recorded diabetes diagnosis. Participants meeting any criterion were classified as having T2D. Detailed outcome classifications are in Additional file 2: Table S5. Statistical analyses Multiple imputation of missing covariates and UPF/MPF intake Before performing inferential analyses, missing data for covariates, UPF/MPF intake, and metabolites were imputed using the random forest-based algorithm implemented in the R package ‘missForest’ (maximum 50 iterations), chosen for its ability to capture complex nonlinear relationships without parametric assumptions [37]. For participants with dietary data at both phases 3 and 5, average intakes were calculated to represent long-term exposure; when data were missing at one phase (8.6% at phase 3; 16.3% at phase 5), imputation enabled calculation of average intake. Among covariates, educational attainment had the highest proportion of missing data (24.3%). Imputed and complete data were compared in Additional file 2: Table S6. Characteristics of the study participants Baseline characteristics of the participants according to tertiles of average UPF and MPF intake from phases 3 (1991–1994) and 5 (1997–1999) were summarised. Continuous variables were summarised as medians (IQR) because several showed skewed, non-normal distributions. Identification of UPF, MPF-related metabolites Metabolites, UPF intake, and MPF intake were log-transformed to reduce skewness and subsequently standardised to z-scores (mean = 0, SD = 1) to enable comparison of effect sizes [38]. Associations between average UPF/MPF intake (phases 3 and 5) and plasma metabolites measured at phase 5 were examined using multivariable linear regression, adjusting for sociodemographic factors, lifestyle variables, total energy intake, dietary quality, and clinical measures (Additional file 2: Table S7–S8). Multiple testing was controlled using the Benjamini–Hochberg procedure to limit false-positive findings while maintaining power in this high-dimensional setting (false discovery rate [FDR]-adjusted p < 0.002). Elastic net regression was then applied to the FDR-significant metabolites to derive the metabolomic signature, without covariate adjustment. Participants were randomly split into a 70% training set (n = 4207) and a 30% testing set (n = 1803); characteristics of the training and testing sets were similar (Additional file 2: Table S9). The same analytic pipeline was applied separately for UPF and MPF intake. The elastic net model involves two hyperparameters: α, which controls the balance between lasso (α = 1) and ridge (α = 0) penalties, and λ, which determines the overall strength of regularisation [39]. Both hyperparameters were tuned on the training set by 10-fold cross-validation: α was optimised using the ‘caret’ package, and λ was then selected as the value minimising cross-validation error (λ.min) using the ‘glmnet’ package. For the UPF signature, the final model selected α = 0.53 and λ = 0.24; for the MPF signature, α = 0.20 and λ = 0.36. The tuned model was refit in the full training set, and model robustness was subsequently assessed using leave-one-out cross-validation within the training set [40]. Metabolites with non-zero coefficients were retained, and the training-set coefficients were used as weights to compute the metabolomic signature score in the testing set (weighted sum). Partial Spearman correlations (covariate-adjusted) were used to relate dietary intakes to metabolomic signature scores. To identify food contributors, covariate-adjusted partial Spearman correlations were calculated between signature metabolites and UPF or MPF food subgroups. Multiple testing was controlled using the more conservative Bonferroni correction, appropriate for this targeted follow-up analysis where minimising false-positive findings was prioritised over discovery, with statistical significance defined as P < 0.05 divided by the number of tests. Survival analysis of UPF and MPF intake, metabolomic signatures, and cardiometabolic outcomes Cox proportional hazards regression models assessed associations between UPF and MPF intake and incident CVD and T2D. Both dietary intakes and metabolomic signature scores were modelled as tertiles and as continuous z-standardised variables (per 1-SD increase). Metabolomic signature scores were calculated as the weighted sum of metabolite concentrations selected by elastic net regression, with weights corresponding to the elastic net coefficients. Follow-up time was calculated from the date of the baseline assessment (phase 5) to the date of the first event (CVD or T2D), death, or the end of the follow-up period, whichever occurred first. Hazard ratios (HR) and 95% confidence intervals (CI) were estimated for each tertile, with tertile 1 (lowest intake) as the reference category. Proportional hazards assumptions were verified using Schoenfeld residuals tests; no violations across all models were detected (all p > 0.05). Two sequential models were fitted. Model 1 adjusted for sociodemographic factors (sex, age, ethnicity, employment grade, marital status, education), lifestyle variables (physical activity, alcohol consumption, smoking status), total energy intake, clinical measures (BMI, dyslipidaemia, hypertension) and family history of CVD and T2D. Model 2 additionally adjusted for dietary quality indicators (sugar, sodium, fat and fibre intake) to examine whether associations were independent of nutrient composition. P for trend was calculated by treating the categorical variable as continuous using the median value within each category. Mediation analysis To quantify mediation, we conducted causal mediation analyses using the R package “mediation”. UPF or MPF intake was specified as the exposure, with incident CVD and T2D as the outcomes. We examined two sets of mediators: (i) the corresponding metabolomic signature score, and (ii) individual nutrient mediators (sugar, sodium, and saturated fat), each modelled separately. For each mediator, linear regression was used for the exposure–mediator relationship and we fitted parametric survival models for time-to-event (accelerated failure time models) including both the exposure and the mediator, adjusting for covariates as specified in Model 1. Dietary intake variables (total sugar, total fat, and sodium) were excluded from mediation models as they may lie on the causal pathway. The treatment and control values for the z-scored exposure were set at 1 and 0. Natural indirect effects (average causal mediation effects), natural direct effects, and total effects were estimated, with 95% CIs obtained via 1000 nonparametric bootstrap resamples. The proportion mediated was calculated from the indirect and total effects and reported only where the indirect effect was statistically significant. These analyses assume no unmeasured confounding of the exposure–outcome, exposure–mediator, and mediator–outcome relationships, and that no mediator–outcome confounder is affected by the exposure. The plausibility of these assumptions was considered in the interpretation of results [41]. Subgroup and sensitivity analyses Stratified analyses were conducted according to sex (male, female), BMI (<25, 25–29.9, ≥30 kg/m²), and smoking status (current, former, and never smokers). Interactions between stratified factors and UPF/MPF intake or the metabolomic signature on the risk of CVD and T2D were tested. To explore which types of UPF drove the observed associations, nine UPF subgroups were each modelled separately as continuous variables (per 1-SD increment, %g/day) in relation to incident CVD and T2D. Several sensitivity analyses were conducted to assess the robustness of findings. First, to assess sensitivity to the choice of imputation method, multiple imputation by chained equations (MICE) was performed, and Cox regression estimates from ten imputed datasets were pooled under Rubin’s rules [42]. Second, complete-case analyses were performed, restricting to participants with no missing data for exposures, covariates, or metabolites (N = 4551), to evaluate potential bias introduced by imputation. Third, lag analyses were conducted, excluding incident cases occurring within the first two years of follow-up, to minimise the potential for reverse causation whereby subclinical disease may have influenced dietary intake or metabolite levels. Lag analyses were applied to: (i) associations between UPF/MPF intake and incident CVD; (ii) associations between metabolomic signature scores and incident CVD; and (iii) corresponding models for incident T2D. Fourth, we adjusted for the Alternative Healthy Eating Index-2010 (AHEI-2010) in place of individual nutrient intake indicators, to examine whether overall diet quality, rather than specific nutrients, accounted for the observed associations. Fifth, to evaluate sensitivity to Nova classification decisions, two alternative reclassification scenarios were constructed for 35 FFQ items whose Nova assignment was potentially ambiguous: an upper bound scenario, in which ambiguous items were reclassified to maximise UPF intake, and a lower bound scenario, in which ambiguous items were reclassified to minimise UPF intake. Spearman correlations and Cox models were estimated under each scenario. Sixth, models examining the association between metabolomic signature scores and incident CVD and T2D were further adjusted for UPF or MPF intake, to assess whether the metabolomic signature captured information beyond self-reported dietary exposure. Seventh, the signatures were re-derived with confounders entered as unpenalised terms in the elastic net, and their correlation with the primary signatures assessed. Eighth, because NMR metabolomic profiling was undertaken at phase 5, analyses were repeated using phase 5 dietary intake alone, rather than averaging intake across phases 3 and 5, to assess whether the findings were robust when dietary exposure was temporally aligned with metabolomic assessment. Finally, to compare the relative discriminative performance of self-reported dietary intake and metabolomic signatures for incident CVD and T2D, Harrell’s C-index was estimated from Cox proportional hazards models with sequential addition of exposures to the base covariate set. Analyses used R 4.4.1 and STATA 18.0 (StataCorp, College Station, TX); significance was set at p < 0.05 (two-sided), unless otherwise specified. Results Baseline characteristics of the 6010 participants free of CVD and T2D at baseline are reported in Table 1. The majority were white (>90%), male (~77%), and married (~76%), with a median age of 55 years. Median overall UPF and MPF intakes were 18.3% (IQR: 14.0–23.5) and 62.1% (54.0–69.3) of g/day, respectively. Across increasing tertiles of UPF intake, median intake rose from 10.1% (10.2–12.3) to 17.5% (16.8–19.9) to 28.6% (23.5–30.5); corresponding medians for MPF intake were 48.0% (44.1–54.0), 62.2% (59.9–64.4), and 71.8% (69.3–75.5). Participants in the highest UPF tertile were more likely to be physically inactive (41.0 vs 32.1%) and to be current smokers (10.3 vs 8.2%), though heavy alcohol consumption was lower in this group (32.3 vs 37.2%). Individuals in the highest MPF tertile were less likely to be current smokers (8.1 vs 11.0%) or to engage in heavy alcohol intake (29.7 vs 36.6%). Table 1Baseline characteristics of study participants by tertiles of ultra-processed food (UPF) and minimally processed food (MPF) intake (N = 6010)VariablesUPF intakeMPF intakeTertile 1Tertile 2Tertile 3Tertile 1Tertile 2Tertile 3Overall Nova food intake (% g/day), median (IQR)a18.3 (14.0–23.5)62.1 (54.0–69.3)Nova food intake (% g/day) by tertiles, median (IQR)a10.1 (10.2–12.3)17.5 (16.8–19.9)28.6 (23.5–30.5)48.0 (44.1–54.0)62.2 (59.9–64.4)71.8 (69.3–75.5) Sex, n (%) Male1503 (75.0)1544 (77.1)1556 (77.6)1589 (79.3)1498 (74.8)1516 (75.6) Female500 (25.0)459 (22.9)448 (22.4)414 (20.7)505 (25.2)488 (24.4) Ethnicity, n (%) White1806 (90.2)1821 (90.9)1830 (91.3)1827 (91.2)1810 (90.4)1820 (90.8) Non-white197 (9.8)182 (9.1)174 (8.7)176 (8.8)193 (9.6)184 (9.2)Age, median (IQR), years54.9 (50.6–61.0)55.1 (50.8–60.9)55.5 (50.7–61.8)55.2 (50.6–61.5)55.1 (50.8–60.9)55.3 (50.6–60.9) Family history, n (%) CVD Yes201 (10.0)204 (10.2)194 (9.7)199 (9.9)202 (10.1)198 (9.9) No1802 (90.0)1799 (89.8)1810 (90.3)1804 (90.2)1801 (89.9)1806 (90.1) T2D Yes159 (7.9)175 (8.7)170 (8.5)126 (6.3)156 (7.8)222 (11.1) No1844 (92.1)1828 (91.3)1834 (91.5)1877 (93.7)1847 (92.2)1782 (88.9) Marital status, n (%) Yes1510 (75.4)1544 (77.1)1520 (75.8)1527 (76.2)1527 (76.2)1520 (75.8) No493 (24.6)459 (22.9)484 (24.2)476 (23.8)476 (23.8)484 (24.2) Smoking, n (%) Current smoker164 (8.2)182 (9.1)206 (10.3)220 (11.0)170 (8.5)162 (8.1) Former smoker779 (38.9)783 (39.1)686 (34.2)757 (37.8)741 (37.0)750 (37.4) Never smoker1060 (52.9)1038 (51.8)1112 (55.5)1026 (51.2)1092 (54.5)1092 (54.5) Alcohol consumption, n (%) Abstainers337 (16.8)304 (15.2)370 (18.5)351 (17.5)294 (14.7)366 (18.3) Moderate drinkers921 (46.0)977 (48.8)986 (49.2)919 (45.9)922 (46.0)1043 (52.0) Heavy drinkers745 (37.2)722 (36.0)648 (32.3)733 (36.6)787 (39.3)595 (29.7) Education, n (%)